ECMWF’s new AI model delivers ten-day forecasts in minutes on a single GPU, replacing equations with learned patterns to predict the atmosphere from memory.
It still feels impossible that the sky can be computed in three minutes.
Yet at the European Centre for Medium-Range Weather Forecasts (ECMWF), a single GPU now renders ten days of weather without solving a single equation.
The system, called AIFS, the Artificial Intelligence Forecasting System, marks the moment when forecasting stopped calculating and began remembering. This is a big deal when it comes to efficiency, accuracy, and scalability for an area that has historically relied on massive number crunching on some of the world’s most powerful supercomputers.
Unlike previous methods of forecasting, AIFS doesn’t simulate air and pressure in motion, but recalls them through patterns learned from decades of atmospheric data. At ECMWF, that shift placed AI beside one of the world’s most refined numerical models.
For decades, weather forecasting relied on physics to advance itself, with each generation of systems adding finer resolution to the equations that shaped the atmosphere on screen. At ECMWF that legacy still drives the numerical system known as IFS, but beside it now runs a model born from data instead of just, well, theory.
AIFS was trained to recognize the atmosphere’s own behavior, learning from patterns across decades of reanalysis instead of calculating them from first principles.
AIFS began as a mere experiment. The first runs, trained on the ERA5 reanalysis archive, produced fast but slightly alien forecasts. They were realistic from a distance but off in key ways. As the authors of the AIFS work note, rain went negative, convective totals exceeded the storms they came from, extremes were smoothed into statistical politeness. And while the researchers could have dismissed it as clever mimicry that couldn’t be counted on, they started to see that if AI could learn the weather, maybe it could also learn the limits of nature.
That insight shaped what became AIFS 1.1.0, released in August 2025.
Physical bounds were written directly into the network’s structure and as of now, the final layers now enforce realism (no negative rainfall, no over-saturated skies, no convective values greater than total precipitation). The model learns within those constraints and the result is both faster and more skilled.
Using AIFS, forecast accuracy improved by about a day over the previous version and over ECMWF’s own numerical system, the IFS.
The model was trained on the GPU-laden Leonardo supercomputer in Italy, part of the EuroHPC program and hosted by Cineca, using mixed precision (which the center has been experimenting with before others) across multiple nodes.
The full training cycle took about three days. Additional compute access came from the MareNostrum 5 supercomputer, the EuroHPC system at the Barcelona Supercomputing Center, awarded through PRACE and the EuroHPC Joint Undertaking.
AIFS stands among a rising field of data-driven forecasters like DeepMind’s GraphCast, Huawei’s Pangu-Weather, China’s FuXi, and FourCastNet from NVIDIA and Microsoft. The difference is that AIFS is not a demo or concept. It’s actually running in production, four times daily, as part of the world’s most respected operational forecasting chain.
Its architecture is based on: a graph neural network that almost “encodes” the atmosphere. A transformer interprets it through attention windows, and the decoder writes it back into physical fields.
Each six-hour prediction becomes the seed for the next. Within minutes, the model produces ten days of global forecast.
What emerged is a system that forecasts with memory instead of mechanics.
At ECMWF, the numerical and neural forecasts now fully coexist in production. IFS still handles the ensemble spread and long-range coupling, but it’s really AIFS providing the instant, high-skill preview. Together they form a hybrid view the atmosphere, one more analytical and the other one learned but what’s cool here is that they continually refine each other.
The researchers also explain that the newest AIFS also sees more of the planet: soil moisture and temperature, runoff, snowfall, solar radiation, 100-meter winds, and multi-layered cloud fields. All of this runs within ECMWF’s Anemoi framework, which was designed for data-driven Earth system modeling that records every checkpoint and dataset for reproducibility.
In that architecture lies the prototype of an Earth-scale foundation model, one capable of learning and bringing together atmosphere, land, and ocean data.
Ten days in three minutes is simply the result of a system that has learned how the atmosphere behaves.
